# Spektr Cold-Read — Full Report # headless.design/portfolio/spektr/llms-full.txt # Gustaf Garnow · 2026-05-14 --- ## the data i ran `spektr.com` through agent.opensverige.se — an open-source scanner i co-built. it scores public surfaces on how legible they are to LLM agents, MCP clients, and discovery protocols. **result: 6 / 14. PARTIALLY READY.** 1 critical issue. 5 warnings. 6 ok. scan permalink → `agent.opensverige.se/scan/spektr.com` ### what's already working - robots.txt allows AI agents - sitemap present - public OpenAPI spec discoverable - proper meta and structured data on key pages - the AI summary buttons in the hero (ChatGPT / Gemini / Claude) — these are genuinely ahead of the curve ### critical - **no markdown content negotiation** — agents fetching spektr.com get full HTML. ~80% wasted tokens vs. serving markdown when `Accept: text/markdown` is set. cloudflare implemented this in their own scanner april 2026. ### warnings - **no `/llms.txt`** — agents can't get a structured brief of what spektr offers. they scrape and infer. - **no `/llms-full.txt`** — no markdown-concatenation of the site. anthropic, cloudflare, and stripe all ship this. - **cookie handling for non-human clients is unclear** — agents fetching the site hit a consent layer that's not declared for non-human user agents. - **AI mentioned but no Article 22 framing** — for a compliance company selling to EU financials, the right to human intervention in automated decisions deserves named placement. ### recommended (not blocking) - **no `/.well-known/mcp`** (SEP-1960) — MCP discovery manifest. Cloudflare added theirs in April 2026. - **no MCP `server-card.json`** (SEP-1649) — capabilities aren't programmatically discoverable. The scanner itself notes: "REST API + OpenAPI are agent-ready without MCP. MCP gives direct tool-coupling for Claude, Cursor, and Windsurf — a good next step but not a blocking requirement." --- ## what this looks like in practice three scenarios where the gap matters today: **1. a compliance officer asks Claude "what does spektr do?"** Claude fetches spektr.com, parses ~3000 lines of HTML, summarizes what it can. the result is shallow and partly invented. with a `/llms.txt`, Claude returns a precise, vendor-controlled summary — every time. **2. an agent-mediated procurement check at a Tier 1 bank.** the agent compares vendors. it pings `/.well-known/mcp` on each. spektr returns 404. that doesn't disqualify, but it places spektr in the "needs manual review" bucket while a competitor with a server-card moves to "auto-eligible." **3. an integration team at Pleo onboards spektr.** they find the OpenAPI (good). but the rest of the docs require scraping. an `/llms-full.txt` cuts the integration scoping from days to hours. --- ## what i'd ship in week one these are additions to the existing surface, not redesigns. minimal risk. maximum agent-leverage. | # | what | priority | time | |---|------|----------|------| | 1 | `/llms.txt` | Important | afternoon | | 2 | `/llms-full.txt` | Important | one day | | 3 | Markdown content negotiation (CDN-level) | **Critical** | one week | | 4 | `/.well-known/mcp` + `server-card.json` scaffolding | bonus | one hour | | 5 | Cookie/consent declaration for non-human user agents | warning fix | half day | | 6 | Article 22 framing on marketing surface | warning fix | afternoon | total: ≈ one week of focused work. zero redesign. --- ## why this is a designer's problem the job post says: *"Designing for AI: This is the frontier."* frontier isn't dashboards. frontier is: - how an LLM **perceives** spektr when it has 8000 tokens of context. - how a procurement agent **decides** whether spektr is in the consideration set at all. - how an MCP client **decides** whether to auto-configure against your tools. - how a compliance officer's assistant **explains** "spektr" when asked. each is a design surface. structure, hierarchy, information architecture, naming, capability articulation. classical design problems, in a new medium. the post asks for *"systems thinker"* and *"AI enthusiast."* designing the agent-readable surface of your own product is the systems-AI intersection — in one project. --- ## what i bring - **10 years in design** — full stack. print, digital, motion, product. 7 of those years at AMF Pension (Sweden's largest pension provider). compliance-adjacent product is not new to me. - **AI-native build practice** — Claude, Cursor, MCP-servers, agent-pipelines, public APIs as part of normal work. not a side project. - **solo operator** — i ship in days, not quarters. - **i built the scanner that scored spektr.com 6/14.** agent.opensverige.se is my work. i'd also build what brings the score to 14. --- ## the move your design works. your product is impressive. your funding is real. the OpenAPI is there — that's already half the battle, and most B2B SaaS in this category hasn't shipped it. what's not yet built is the protocol-layer surface beneath the marketing site. that's the frontier the role names. that's the work i want to do. easy apply submitted. CV attached. this page is the work sample. — gustaf --- **find me** - gustaf.garnow@gmail.com - https://headless.design - https://linkedin.com/in/gustafgarnow - https://github.com/Baltsar - discord @baltikum - currently Stockholm · Skåne if this lands — Øresund commute is 35 min